AI Trade Signals
The Scalability of AI for Execution
AI Trade Signals
Order execution is the process of completing a buy or sell order in the market. While many investors believe this is an instantaneous process — perhaps automated, many are unaware that there is an actual trader or dealer who trades the stocks on their behalf. When trading low volume, investors will have a high chance to buy and/or sell at the designated market price. With high volume trades, however, a lot of effort will be required by the traders to reduce the transaction cost.
At the heart of the financial industry is the traders. They represent the fierceness of Wall Street and the underlying psychology that drives market-oriented investments. They are intuitive, fast-responding, and pragmatic with their order execution. But in recent years, the situation surrounding traders has been unusual.
It all started when Goldman Sachs reduced the number of traders from 500 people to just merely three and announced that those traders were to be replaced by robots. (Bloomberg, A Goldman Trading Desk That Once Had 500 People Is Down to Three)
JP Morgan also announced a deep reinforcement learning based equity execution system called LOXM to be tested on the European stock markets.
Similar news has also been appearing in South Korea.
Order Execution Competition: AXE vs. Human Traders
Order Execution Competition: AXE vs. Human Traders
Rule-Based Order Execution System
While AI-driven order execution systems are being newly released — thanks in large part to the recent developments in the field of Artificial Intelligence, automated rule-based order execution systems based on machine learning algorithms have been around for quite some time.
For example, the simplest algorithm, TWAP — also known as Time Weighted Average Price Algorithm, is the average price of a security over a certain period. Specifically, it divides and orders a large quantity mechanically by 1/N for a given time. In addition, a more complex VWAP algorithm predicts the trading volume curve for the day and trades on the hypothesis that daily trading volume is stable. In essence, if the trading volume is high, you buy a lot. And if the volume is low, you buy less. Other rule-based order execution algorithms include IS, PoV, and IMSH. In addition, there are many companies like ITG (which is recently acquired by Virtu Financial), that provide order execution services based on these algorithms.
The United States is not only home to the largest stock exchange in the world, but it also credits the largest transaction volume compared to other international markets since there are no transaction taxes involved. And a significant portion of these transactions is credited by the algorithmic order executions precisely because of their reduced costs.
If you buy and sell thousands and millions worth of stocks at market prices without splitting them, you are more inclined to make a bigger market impact which in turn will increase your transaction costs. Therefore, large orders are sliced into many small orders in execution. However, if traders use predictable algorithms to split the orders, they are more likely to fall prey to other investors who will take advantage of the system. It is changed now, but at one point the main source of income for high-frequency trading companies was to buy and sell the stocks first by detecting large installment orders of mutual funds early. This has led to an increase in transaction costs for mutual fund companies and the order execution algorithm has since changed to more complex systems so that other traders will not notice easily.
AI Order Execution (algo trading)
Fortunately, AI order execution is one of the few areas in the field of Artificial Intelligence that can be commercialized promptly and see immediate performance results. In other words, AI order execution is not just a replica of a rule-based order execution system with the fancy “AI” words attached.
Contrary to what most people believe, the AI order execution system does not simply predict whether the market will go up or down based on learning market flows through AI techniques. Instead, the system has a structure that resembles the high-frequency trading system.
While it is almost impossible to predict the daily stock market or the movement of individual stocks, there are several short-term patterns that are somewhat significant when entering the tick data of individual stocks, including the bid/ask price data. For example, TESLA is more likely to see its price rise if the number of buy orders increases when there is a substantial amount of orders in the ask price (not true, just arbitrary example). In particular, in South Korea, 0.2% of transaction taxes exist, so even if a particular short-term trading strategy is effective if the yield per transaction is less than 0.2%, it is almost impossible to make money from that strategy. And keep in mind, these short-term patterns often have a very long lifespan. In addition, the patterns differ for each stock and will continue to change in the long run.
The AI order execution system learns the short-term patterns of these individual stocks through deep reinforcement learning to find the optimal order execution strategy. Based on the South Korean market, stock tick data is generated daily in units of several gigabytes, which is much larger than other financial sectors like asset management. This in turn better demonstrates the full spectrum of AI’s learning performance.
On the other hand, even if an individual stock tick data is properly learned through deep reinforcement learning and an appropriate order execution strategy is used, the predicted success rate of divided individual orders is about 50–60%. However, with more than 500–700 orders being divided and executed, the probability of beating VWAP on a daily basis is actually 80–90%. (Think of it this way — if you play a game that has a 55% winning rate over 500 times, you will most likely win the money.) In the same way, if you execute a portfolio with diversified assets, your odds of positive returns will most likely increase.
AXE has the power to process infinite numbers so long as the computing power increases. In the case of human traders, experienced veterans can handle at most 30 shares per day. And since human traders cannot analyze tick data for individual shares, most of them handle it with intuition and experience. (For example, when dealing with rule-based orders like VWAP, human traders will be very busy just to calculate the VWAP curve.)
It will take at least several years to confidently claim that AI performs better at managing assets than humans in the field of asset management. However, with regards to execution orders, the difference in performance can be seen in just a few short days. This is because there is almost no chance of winning by luck because each order is divided into 500–700 sliced orders a day. (Law of large numbers) Human traders are almost impossible to beat VWAP, which is the benchmark for order execution and the average execution performance of a human trader is known as 5–10 bps inferior to VWAP. But, AXE applied to financial institutions consistently show superior performance than VWAP through real market execution of over $300m/month.
Because of the learning-based structure, AI order execution systems are designed to learn tick data daily or periodically. JP Morgan’s LOXM system, the first in this field, also works by learning stock tick data with a deep reinforcement learning model.
JP Morgan’s LOXM
Architecture of LOXM
LOXM is a system that is designed to solve multiple-choice questions.
According to JP Morgan’s presentation at the NIPS workshop, LOXM uses a PoV (Percentage of Volume, a simple algorithm that keeps buying a certain percentage of the transaction volume) rule-based algorithm. The market order is not used (However, the lower limit is set and the transaction is not made too much, the market order is forcibly placed). LOXM only put a limit order in the bid prices, but LOXM can also select in three options of the amount when placing an order: [small, medium, and large]. While using the PoV algorithm, the order is allocated only to the first bid price, the second bid price, and the third bid price.
#1 [bid1: small, bid2: small, bid3: small]
#2 [bid1: medium, bid2: small, bid3: small]
#3 [bid1: large, bid2: small, bid3: small]
#4 [bid1: small, bid2: medium, bid3: small]
…
#27 [bid1: large, bid2: large, bid3: large]
By deep reinforcement learning through matching 27 options with individual tick data, AI can deduce which of the 27 options fit the best.
According to JP Morgan, LOXM’s future goal is to increase the number of action spaces (currently composed of multiple choices), and ultimately determine the amount and type of orders directly (not just limit orders as they are known today, but whether to use market orders if necessary).
Fortunately, JP Morgan doesn’t seem to need this type of research. That’s because Qraft Technologies’ AXE system has already integrated a similar approach and since proven its performance in the real market. Of course, when JP Morgan first developed LOXM, the AI technology was at the level of DQN. In the years between LOXM and AXE, deep reinforcement learning technology has developed remarkably and Qraft Technologies — while late to the game, has been able to produce substantial results ever since.
Qraft Technologies’ AXE
Qraft’s AXE uses a state-of-the-art agent-based deep reinforcement learning model to place a certain amount of orders at the different bid/ask prices and even at market price and calculates with end-to-end AI to even correct and cancel the orders. The structure of AXE is as follows — note that the architecture below is the previous version of the current AXE. The latest AXE model, which has been enhanced, will be applied to the upcoming API version.
AI Model applied to Qraft Technologies’ AXE
Real Trading Performance of NPS orders by AXE: 82.5% of the whole transaction outperforms top dealer (National Pension Fund level 1 dealer)
At the end of 2018, NVIDIA, Shinhan Bank, KOSCOM, and PwC sponsored a competition called the AXE Challenge with a total grand prize of $100,000 dollars. The challenge was essentially for human traders from well-known securities firms to compete against AXE in buying a portfolio at a cheaper price. The portfolio was randomly generated by KOSCOM. The competition lasted for five days and was broadcasted live on TV. At the end, the AXE system had beat the human traders by a large margin.
Since March 2020, the Shinhan Financial Group has been actively incorporating AXE in their order execution system. So far, it has been able to process about over $100 million dollars in stock orders each month. AXE has executed equity orders from the National Pension Service and showed 2~3bps outperformance when compared to VWAP.
The AI order execution system explores the most optimal strategy by learning patterns from tick data. This includes not only the price and transaction volume of individual stocks but also transaction details and limit order book data. When such a system is introduced, it is possible to improve the performance by minimizing the transaction costs of mass trading of all financial products as well as the active index. In particular, the effect is great in large funds with a big market impact, funds targeting small and medium-cap stocks, and funds with a high turnover rate.
AI Execution in B2C
Recently, AI order execution has taken a slightly unexpected direction. Up until now, the order execution system has been a topic of interest to only the corporate sales and trading department and asset management companies as well as securities firms. This is mostly due to the demanding nature of their work and how technology could add convenience to their handling orders. However, with the recent entry for fintech platforms in securities firms, a whole new market for AI order execution has emerged.
Take for instance, a messenger platform that wants to incorporate stock trading or asset management features into the app. (In fact, while not a messenger platform, companies like GRAB have already added asset management function into the app and have plans to include stock trading functions as well.) The value that many of these app-based platforms pursue is UI/UX. Currently, the demand for every consumer in any field is simplicity. But with stocks, the concepts are already too difficult to start with. This is especially true for mobile-based platforms that are consumer-focused, not unlike many of the securities companies. For example, when you select a stock, the price keeps moving, the buy and sell quantity are also changing. In most cases, you don’t know where to place your orders because the market keeps shifting.
With the implementation of AI order execution technology, messenger platforms can simplify the complex UI system for stock trading and create a tremendously easy and intuitive service with just mere voice or words. To put it simply, with messenger platforms, you can simply just say or type “Please buy me $1,000 dollars of Tesla Stocks till next week” and the process will be finished. No longer will investors have to worry about when to buy or sell while looking through MTS. If this type of UI/UX becomes widespread, complex MTS that requires order price, market price/specified price selection, etc. will become a relic of the past.
Soon, limit order book screen and the complicated ordering interface will disappear.
Example: Conversational Trading with AXE
Example: 4 Button Trading UI with AXE, No Limit Order Book, Just 4 Button for full trading service
Robinhood, famous for its free stock trading service, also offers a rudimentary order execution service within the app. For example, if a specific stock is bought or sold somewhere between $5 to $6 dollars, the process becomes automatic once the stock reaches the corresponding price. It is a feature that’s rarely used because of its lack of certainty of order completion. Nevertheless, the reason Robinhood had no choice but to make it like this is that the market order is surely completed, but it is too risky because we do not know what the bid-ask and price situation will be. And limit order is impossible to know whether order will be completed or not. As a result, Robinhood, known for its intuitive UI/UX, also compromised in the order execution function to the extent that the customer itself sets up the price band.
AI Execution can complete transactions by the given deadline 100% of the time while using the best order execution strategy to achieve a good price.
AI Execution as a Service
Qraft Technologies recently announced that it will package the AI order execution system AXE in the form of a cloud-based API with Microsoft. (Surely, we will expand it to AWS, too) This means that by the end of the year, once the cloud solution is finalized, anyone can apply Qraft’s AXE to their services through cloud subscriptions and API connections. (The server version of AXE, which was supplied to the Shinhan Financial Group through the ‘on-premise’ type, will be discontinued. The new AXE API will be provided as a model to charge several bps of the transaction volume.)
In order to automate stock trading, an automated order execution function is absolutely essential. However, the rule-based order execution system is difficult to trust because of its low performance and the lack of transaction completeness (for example, 100 shares must be finalized if you order 100 shares). That’s why AI execution is the only alternative out there after all.
AI order execution requires a lot of computing power compared to simple rule-based systems. However, the computing power required in the learning process is a fixed cost irrespective of the increase in the number of customers. The computing process required in the inference process, which increases variably with the increase in the number of customers, is much smaller than that of the learning process, enabling cost management below 10% of the algorithm’s sales. This is due to the GPU performance that’s rapidly improving.
NVIDIA, the world’s largest GPU company, recently selected Qraft Technologies among 30 AI startups worldwide for its inception premier program that supports marketing and sales. Qraft is the first startup in South Korea to be chosen and the only one in the world from the financial sector. (Other selected companies are mostly unicorn startups or famous AI firms that have been acquired by Apple and other blue-chip companies.) NVIDIA’s selection of Qraft Technologies in the premier program may be a sign of anticipation that if AXE penetrates the global financial market, NVIDIA’s GPU will also be sold substantially. With AI order execution, the more customers order, the more real-time simultaneous processing capability is required. And AXE requires a lot of parallel GPU computing power in the learning process and inference process.
Customers of domestic and global financial institutions or platform companies will get consulting on AI transformation from IBM, PwC, etc.
IBM and PwC apply Qraft’s API in the process of AI transform project, and their customers will use AWS cloud or Azure cloud to use Qraft’s AXE API. It is a value chain that customer subscribes to the corresponding cloud instance for AXE API and sales of NVIDIA GPUs will increase.
Inception Premier program introduces AXE to NVIDIA’s financial institution clients through NVIDIA’s regional marketing/sales manager, supports related seminars and roadshows, and further conducts co-marketing and co-sales.
Future of Execution
AI order execution is still in its nascent stage. And it’s a rare field where AI technology can make real and direct changes quickly, has large-scale commercial potential, and is a relatively easy market to penetrate.
By applying AI technology, it’s now possible to catch two rabbits: the order execution’s completeness and excellent performance. In addition, proof of that performance is easy to see.
This has made it possible to fully automate and abstract functions related to stock trading on rapidly growing platforms and various services.
This change also makes it easy for investors to invest and enable intuitive and easy investment services.
Pretty soon, investors placing orders or trading themselves will look very similar to driving a sports car for fun in the era of autonomous driving.
Disclaimer
The past performance may not be indicative of future results.
This material was prepared for informational purposes and cannot be used for the purpose of soliciting the sale of financial investment products such as funds.
This document contains the contents of the patent-pending or registered by Qraft Technologies, Inc.